Download Homework 1 - Emerson Statistics Home

Biost 524, Spring 2011
Homework #1
March 30, 2011, Page 1 of 5
Biost 524: Design of Medical Studies
May/Emerson, Spring 2011
Homework #1
March 30, 2011
Written problems due at the beginning of class, Monday, April 4, 2011.
This homework explores the role of screening trials in promoting the accuracy of the process.
The goal of the drug approval process should be
1. To have a low probability of approving drugs that do not work,
2. To have a high probability of approving drugs that do work, and
3. To have a high probability that an approved drug does work.
Now suppose we decide to perform a experiment or series of experiments, and to approve the drug
whenever the estimated treatment effect (perhaps standardized to some Z score) exceeds a predefined threshold. When stated in statistical jargon, these goals become
1. To have a low type I error  when a null hypothesis of no treatment effect is true,
2. To have a high statistical power Pwr= 1- (so  is the type II error) when some alternative
hypothesis is true, and
3. To have a high positive predictive value PPV = (number of approved effective drugs) /
(number of approved drugs).
We can examine the interrelationships of these statistical design criteria in the context of a RCT
where we let θ denote our treatment effect, and we presume that an ineffective drug has θ = 0, and an
effective drug has some θ > 0.
In the “frequentist” inference most often used in RCT, we typically choose some value for the “level
of significance” (or type I error) . This will be the probability of approving the drug when θ = 0.
Most often, we base our decisions on some estimate of the treatment effect that is known to be
approximately normally distributed
 V
ˆ ~ N  , .
 n
In experimental design, we sometimes choose a sample size n and then compute the power of the
study to detect a particular alternative hypothesis. When our null hypothesis corresponds to θ = 0,
the power of a particular design depends upon the type I error , the variability of the data V, the
true value of the treatment effect θ, and the sample size n according to the following formula:

n
,
Pwr  1  Pr Z  z1  
(Eq. 1)

V


where Z is a random variable having the standard normal distribution, and the constant z1- is the 1 quantile of the standard normal distribution such that Pr( Z < z1-) = 1 - .
Biost 524, Spring 2011
Homework #1
March 30, 2011, Page 2 of 5
In other settings, we choose a desired power Pwr = 1 - , and then compute a sample size according
to the value of  using the following formula (which again presumes a null hypothesis of θ = 0):
2

z1  z1   V
(Eq. 2)
n
,
2

where we again use the quantiles of the standard normal distribution. The following table provides
values of z1- for selected values of :

z1-
0.005
2.575829
0.01
2.326348
0.025
1.959964
0.05
1.644854
0.10
1.281552
0.20
0.841621
More generally, we can obtain an arbitrary quantile using statistical software. The commands to
obtain the z1- quantile when  = 0.075 in three commonly used programs are:
 (Stata) di invnorm(1 – 0.075)
 (R)
qnorm(1 – 0.075)
 (Excel) norminv(1 – 0.075, 0 , 1)
Similarly, we can obtain Pr( Z < c) for arbitrary choices of c using statistical software. The
commands to obtain Pr( Z < c) when c = 1.75 in three commonly used programs are:
 (Stata) di norm(1.75)
 (R)
pnorm(1.75)
 (Excel) normdist(1.75, 0 , 1, TRUE)
Bayes Rule can be used to compute the PPV from  and , providing we know the prior probability
 that a treatment would work (this prior probability might be thought of as the proportion of
effective treatments among all treatments that we would consider testing—sort of a prevalence of
good treatments):
1     
(Eq. 3)
PPV 
1         1   
In this homework, we consider a couple examples of two different strategies of testing for
experimental treatments:
1. Strategy 1: Test each treatment in one large “pivotal” RCT.
2. Strategy 2: Test each treatment in one small “pilot” RCT that screens for promising
treatments. Any treatment that passes this screening phase, is then tested more rigorously in
one larger “confirmatory” RCT.
To compare “apples with apples”:
 We pretend that we have 1,000,000 patients with disease X to use when evaluating ideas that
we have formulated for treating disease X.
 We further pretend that 10% of our ideas correspond to drugs that truly work (so  = 0.10),
and all those truly effective drugs provide the same degree of benefit θ = 1 to patients with
disease X. The other 90% of our ideas correspond to drugs that provide no benefit to the
patients (so θ = 0).
 In every RCT, the true variability of the patient data corresponds to V = 63.70335.
Biost 524, Spring 2011
Homework #1
March 30, 2011, Page 3 of 5
Problems using Strategy 1: Only Pivotal RCT
1. (A: Pivotal) Suppose we choose a type I error of  = 0.025 and a power of 97.5% (so  =
0.025) under the alternative hypothesis that the true treatment effect is θ = 1.
a. What sample size n will be used in each RCT?
979
z
n
1
 z1   V
2
2
2

1.959964  1.959964  63.70335

 978.855
12
b. How many of our ideas will we be able to test?
1,021
1,000,000 / 979 = 1021.45
c. How many of those tested ideas will be truly beneficial drugs?
102
1,021 x 0.10 = 102.1
d. How many of the tested beneficial drugs will have significant results?
99
102 x 0.975 = 99.47
e. How many of those tested ideas will be truly ineffective drugs?
919
1,021 – 102 = 919
f. How many of the tested ineffective drugs will have significant results?
23
919 x 0.025 = 22.975
g. How many of the tested drugs will have significant results?
122
99 + 23 = 122
h. What proportion of the drugs with significant results will be truly beneficial? 0.8115
99 / 122 = 0.8115 or
1     
1  0.025  0.10
PPV 

 0.8125
1         1    1  0.025  0.10  0.025  1  0.10
2. (B: Pivotal) Suppose we choose a type I error of  = 0.025 and a power of 80.0% (so  =
0.20) under the alternative hypothesis that the true treatment effect is θ = 1.
a. What sample size n will be used in each RCT?
_______
b. How many of our ideas will we be able to test?
_______
c. How many of those tested ideas will be truly beneficial drugs?
_______
d. How many of the tested beneficial drugs will have significant results?
_______
e. How many of those tested ideas will be truly ineffective drugs?
_______
f. How many of the tested ineffective drugs will have significant results?
_______
g. How many of the tested drugs will have significant results?
_______
h. What proportion of the drugs with significant results will be truly beneficial?_______
Biost 524, Spring 2011
Homework #1
March 30, 2011, Page 4 of 5
3. (C: Pivotal) Suppose we choose a type I error of  = 0.05 and a power of 80.0% (so  = 0.20)
under the alternative hypothesis that the true treatment effect is θ = 1.
a. What sample size n will be used in each RCT?
_______
b. How many of our ideas will we be able to test?
_______
c. How many of those tested ideas will be truly beneficial drugs?
_______
d. How many of the tested beneficial drugs will have significant results?
_______
e. How many of those tested ideas will be truly ineffective drugs?
_______
f. How many of the tested ineffective drugs will have significant results?
_______
g. How many of the tested drugs will have significant results?
_______
h. What proportion of the drugs with significant results will be truly beneficial?_______
Problems using Strategy 2: Screening pilot RCT, followed by Confirmatory RCT
4. (D: Screening pilot study) Suppose we choose a type I error of  = 0.025 and a sample size
of n = 100 for each pilot RCT.
a. Under the alternative hypothesis θ = 1, what is the power?
_______
b. If we use 700,000 patients in pilot RCT, how many ideas will we test?
_______
c. How many of those tested ideas will be truly beneficial drugs?
_______
d. How many of the tested beneficial drugs will have significant results?
_______
e. How many of those tested ideas will be truly ineffective drugs?
_______
f. How many of the tested ineffective drugs will have significant results?
_______
g. How many of the tested drugs will have significant results?
_______
h. What proportion of the drugs with significant results will be truly beneficial?_______
5. (D: Confirmatory trials) Suppose we choose a type I error of  = 0.025 and use all remaining
patients in the confirmatory trials of each drug that had significant results in problem 4.
a. How many confirmatory RCT will be performed?
_______
b. What sample size n will be used in each RCT?
_______
c. Under the alternative hypothesis θ = 1, what is the power?
_______
d. How many confirmatory RCTs will be for truly beneficial drugs?
_______
e. How many of the tested beneficial drugs will have significant results?
_______
f. How many confirmatory RCTs will be for truly ineffective drugs?
_______
g. How many of the tested ineffective drugs will have significant results?
_______
h. How many of the tested drugs will have significant results?
_______
i. What proportion of the drugs with significant results will be truly beneficial?_______
Biost 524, Spring 2011
Homework #1
March 30, 2011, Page 5 of 5
6. (E: Screening pilot study) Suppose we choose a type I error of  = 0.10 and a power of
85.0% (so  = 0.15) under the alternative hypothesis that the true treatment effect is θ = 1.
a. What sample size n will be used in each RCT?
_______
b. If we use 700,000 patients in pilot RCT, how many ideas will we test?
_______
c. How many of those tested ideas will be truly beneficial drugs?
_______
d. How many of the tested beneficial drugs will have significant results?
_______
e. How many of those tested ideas will be truly ineffective drugs?
_______
f. How many of the tested ineffective drugs will have significant results?
_______
g. How many of the tested drugs will have significant results?
_______
h. What proportion of the drugs with significant results will be truly beneficial?_______
7. (E: Confirmatory trials) Suppose we choose a type I error of  = 0.025 and use all remaining
patients in the confirmatory trials of each drug that had significant results in problem 6.
a. How many confirmatory RCT will be performed?
_______
b. What sample size n will be used in each RCT?
_______
c. Under the alternative hypothesis θ = 1, what is the power?
_______
d. How many confirmatory RCTs will be for truly beneficial drugs?
_______
e. How many of the tested beneficial drugs will have significant results?
_______
f. How many confirmatory RCTs will be for truly ineffective drugs?
_______
g. How many of the tested ineffective drugs will have significant results?
_______
h. How many of the tested drugs will have significant results?
_______
i. What proportion of the drugs with significant results will be truly beneficial?_______
Comparisons
8. Of the 5 different strategies considered (problems 1, 2, 3, 4 and 5, or 6 and 7) which do you
think best and why?